What the data for this category looks like
Data related to refined oil product sales service inquiries and responses comes primarily from gas station operation ledgers, oil price tag public documents, member service FAQs, inventory purchase-sales-stock reports, and industry policy documents. Update frequencies vary: oil price and inventory data is updated daily, member service FAQs are updated irregularly, and industry policy documents are updated quarterly. Document types include multi-column structured Excel tables, fixed-format PDF service guides, and Word-format inquiry response templates. Fields cover oil product model, specification, unit price, inventory quantity, and handling process steps. Common units include liters, tons, yuan per liter, yuan per ton, and others.
What constraints these characteristics impose on document parsing and chunking
In FastGPT v4.17.0, structured documents for refined oil sales are mostly multi-column Excel tables where each row corresponds to a single business record. Automatic chunking can mix and split fields from different columns, leading to failed retrieval of complete business information in subsequent steps. Some documents contain mixed units, such as conversion relationships between liters and tons; chunk truncation breaks unit relevance. Large inventory reports have large file sizes, and default upload and parsing limits do not meet these needs. Additionally, in service inquiry scenarios, only the currently uploaded attachment must be parsed, to avoid mixing historical documents into parsing results which reduces response accuracy.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Matches the length of refined oil business paragraphs and single Excel records, avoids truncating units or core fields |
chunkSplitter | Split by row (Excel), split by paragraph (PDF) | Matches splitting rules for different document types, resolves mixed formatting issues in multi-column Excel |
trainingType | Structured document priority | Preserves field integrity of structured tables first, adapts to the fixed-format characteristics of refined oil documents |
UPLOAD_FILE_MAX_SIZE | 500–1000 MB | Meets file size requirements for large inventory reports and bulk uploaded price tag documents |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Provides sufficient parsing time for large Excel documents with multiple sheets, avoids timeout errors |
chunkSettingMode | Custom mode | Supports separate chunking rule configuration for different document types, avoids chaotic results from automatic chunking |
Three Common Mistakes
- Phenomenon: After uploading an Excel file, chunking results mix content from multiple columns or split single business records. Cause: Did not specify row-based chunking rules, used the default paragraph splitting logic.
- Phenomenon: Parsing task results include content from previously uploaded documents. Cause: Did not enable the switch to parse only the currently uploaded file, or retained global historical document loading configuration.
- Phenomenon: Parsing timeout error occurs after uploading a large inventory report. Cause: Did not adjust the
PARSE_FILE_TIMEOUT_SECONDSparameter, used the default shorter timeout duration.
How to Confirm Proper Configuration
- Upload an Excel snippet containing a single business record, check if the chunking result is complete single-row business content.
- After uploading batch documents, view the parsing task list to confirm only currently uploaded files are included, with no historical documents mixed in.
- Upload an inventory report larger than 500 MB, verify that the parsing process does not trigger timeout errors.
- Adjust
chunkSizeto the preset value, randomly sample chunking results to confirm the length meets expected business paragraph requirements.
The values listed are common starting points and should be evaluated against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.